Adverse drug reactions (ADRs) are dangerous and expensive, afflicting about 1.5% of hospitalized patients with profound health and financial consequences. In Medicare patients alone, adverse drug reactions account for 19% of total spending ($339 billion), more than 1,900 deaths, and more than 77,000 extra hospital days per year. Idiosyncratic ADRs, especially rare and severe hypersensitivity-driven ADRs, are the leading cause of medicine withdrawal and termination of clinical development. At the same time, a large proportion of drugs are not associated with hypersensitivity driven ADRs, offering hope that new medicines could avoid them entirely with reliable predictors of risk. Hypersensitivity driven ADRs are caused by the formation of chemically reactive metabolites by metabolic enzymes. These reactive metabolites covalently attach to proteins to become immunogenic and provoke an ADR. Unfortunately, current computational and experimental approaches do not reliably identify drug candidates that form reactive metabolites. These approaches are limited because they inadequately model metabolism, which can both render toxic molecules safe and safe molecules toxic. To overcome this limitation, the proposed study aims to curate a public database of metabolism and reactivity and use this database to build accurate and validated mathematical models of metabolism and reactivity. The models will be constructed using machine-learning algorithms that quantitatively summarize the knowledge from thousands of published studies. The Aims are to (1) curate a database of metabolism and build models that identify rules governing the structure of reaction products during drug metabolism in the liver, (2) curate a database of reactivity and build improved reactivity models that mechanistically predict which metabolites are reactive with biological molecules, and (3) curate a database of reactive metabolites and combine these models to predict when molecules form reactive metabolites that covalently bind proteins. The computational models generated by these Aims will be validated through statistical approaches and against bench-top experiments. Taken together, this approach will substantially improve on existing approaches by more accurately modeling the properties determining whether metabolism renders drugs toxic or safe. The predictive models will make new medicines safer by helping researchers avoid molecules prone to ADRs without harming patients.